Keyphrase Annotation with Graph Co-Ranking

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 2945–2955, Osaka, Japan, December 11-17 2016. Keyphrase Annotation with Graph Co-Ranking Adrien Bougouin and Florian Boudin and B´eatrice Daille Universit´e de Nantes, LINA, France {adrien.bougouin,florian.boudin,beatrice.daille}@univ-nantes.fr Abstract Keyphrase annotation is the task of identifying textual units that represent the main content of a document. Keyphrase annotation is either carried out by extracting the most important phrases from a document, keyphrase extraction, or by assigning entries from a controlled domain-specific vocabulary, keyphrase assignment. Assignment methods are generally more reliable. They pro- vide better-formed keyphrases, as well as keyphrases that do not occur in the document. But they are often silent on the contrary of extraction methods that do not depend on manually built re- sources. This paper proposes a new method to perform both keyphrase extraction and keyphrase assignment in an integrated and mutual reinforcing manner. Experiments have been carried out on datasets covering different domains of humanities and social sciences. They show statisti- cally significant improvements compared to both keyphrase extraction and keyphrase assignment state-of-the art methods. 1 Introduction Keyphrases are words and phrases that give a synoptic picture of what is important within a document. They are useful in many tasks such as document indexing (Gutwin et al., 1999), text categorization (Hulth and Megyesi, 2006) or summarization (Litvak and Last, 2008). However, most documents do not provide keyphrases, and the daily flow of new documents makes the manual keyphrase annotation impractical. As a consequence, automatic keyphrase annotation has received special attention in the NLP community and many methods have been proposed (Hasan and Ng, 2014). The task of automatic keyphrase annotation consists in identifying the main concepts, or topics, ad- dressed in a document. Such task is crucial to access relevant scientific documents that could be useful for researchers. Keyphrase annotation methods fall into two broad categories: keyphrase extraction and keyphrase assignment methods. Keyphrase extraction methods extract the most important words or phrases occurring in a document, while assignment methods provide controlled keyphrases from a domain-specific terminology (controlled vocabulary). The automatic keyphrase annotation task is often reduced to the sole keyphrase extraction task. Unlike assignment methods, extraction methods do not require domain specific controlled vocabularies that are costly to create and to maintain. Furthermore, they are able to identify new concepts that have not been yet recorded in the thesaurus or ontologies. However, extraction methods often output ill-formed or inappropriate keyphrases (Medelyan and Witten, 2008), and they produce only keyphrases that actually occur in the document. Observations made on manually assigned keyphrases from scientific papers of specialized domains show that professional human indexers both extract keyphrases from the content of the document and as- sign keyphrases based on their knowledge of the domain (Liu et al., 2011). Here, we propose an approach that mimics this behaviour and jointly extracts and assigns keyphrases. We use two graph representa- tions, one for the document and one for the specialized domain. Then, we apply a co-ranking algorithm to perform …

Content truncated to 500 words